Deep learning for oncological medical image

Abstract

Abstract Background: The emergence of deep learning (DL) in the field of medical imaging has brought significant improvements, especially in pathology and radiology. In pathology, artificial intelligence (AI) methods can detect microsatellite instability (MSI) or mismatch repair deficiency (dMMR), a key genetic feature that should be tested in every patient with colorectal cancer (CRC), from routine pathology slides. On the other hand, in radiology, AI can also help clinicians to identify and detect lung cancer abnormalities from chest X-ray (CXR) images. Methods: To develop the MSI prediction system, we collected H&E-stained slides and findings from molecular analysis for MSI/dMMR of CRC patients (249 patients from Hospital del Mar and 360 from The Cancer Genome Atlas –TCGA– repository). To develop the lung cancer classification system, we collected CXR images of 1679 patients with non-matched controls. Results: We developed two different types of DL models. The first one to predict MSI/dMMR status from routine H&E wholeslide images (WSI) of CRC patients, achieving performance metrics of AUC up to 0.91, sensitivity of 79.5%, specificity of 82.5% for the TCGA’s cohort, and AUC up to 0.96, sensitivity of 100% and specificity of 82.3% for the subgroup of surgical specimens of the Hospital del Mar’s cohort. The second one, for the classification of lung cancer from CXR images, achieving AUC values over 0.99, AUPRC over 0.99, sensitivity and specificity values up to 99% and accuracy of 97%. Conclusion: The system for MSI prediction from H&E slides of CRC is not yet optimal or robust for a routine clinical application but the results are encouraging. On the other hand, given the excellent performance of the classification model of lung cancer from CXR images, we are ready to move forward to an external validation and the development of a detection model.